神经网络类机理建模下的持续自学习控制

针对未知、时变复杂动力学系统在基于模型的控制中的动态建模问题, 本文采用前向全连接神经网络对动力学系统进行数据驱动下的非机理拟合建模. 通过动态线性化和归一化/反归一化数据处理, 基于前向传播算法, 将神经网络的网络拓扑计算过程转化成动力学系统机理模型的同构等价表达形式. 与基于模型的预测与反演控制相结合, 提出了神经网络类机理建模下的持续自学习控制方法, 探索了神经网络在动力学系统建模与控制中的可解释性问题. 以机械臂为控制对象的仿真结果表明, 神经网络类机理模型与机理模型在形式上同构, 在参数上近似或等价, 可用于控制系统控制品质的定性、定量分析. 持续自学习控制对非线性未知、时变复杂系统具有较好的动态适应能力.

Continuous self-learning control under transformation of neural network into isomorphic equivalent form of mechanism model

Aiming at the problem of dynamic modeling of unknown and time-varying complex dynamic systems in model-based control, a forward fully connected neural network is used to model the dynamic system with data-driven non mechanism fitting. Through dynamic linearization and normalization / anti normalization data processing, based on the forward propagation algorithm, the topological calculation process of neural network is transformed into isomorphic equivalent expression of dynamic system mechanism model. Combined with model-based prediction and inversion con-trol, a continuous self-learning control method based on the neural network mechanism modeling is proposed, and the interpretability of neural network in dynamic system modeling and control is explored. The simulation results with the manipulator as the control object show that the neural network mechanism model is similar to the mechanism model in form, approximate or equivalent in parameters, and can be used for the qualitative and quantitative analysis of the control quality of the control system. Continuous self-learning control has good dynamic adaptability to nonlinear unknown and time-varying complex systems.